A system for the subtyping diagnosis of diabetes

Abstract

The application relates to the technical field of medical time series data disease typing, and discloses a diabetes typing and diagnosing system. The system comprises the following steps: generating patient time series data through data preprocessing, and cutting through a multi-scale sliding window, synchronously calculating statistical features and trend slopes of physiological indexes in each window, and constructing a mode trajectory of individual health evolution. The trajectory is mapped to a high-dimensional space to divide diabetes subtypes. For a new data point, multi-dimensional distances and migration probabilities to each prototype are calculated to generate a preliminary subtype membership vector. The vector is fed back to the trajectory construction process to dynamically adjust window parameters and calculation weights in an iterative manner, and after optimization, a stable personal health trajectory and an updated membership are output. Finally, a structured subtype diagnosis with a confidence level is generated. The system realizes dynamic, accurate and individualized typing of diabetes subtypes.

Description

Technical Field

[0001] This invention relates to the field of medical time-series data disease classification technology, specifically a classification and diagnostic system for diabetes. Background Technology

[0002] Clinical classification of diabetes has long relied on single or limited-number measurements of static indicators such as fasting blood glucose and glycated hemoglobin. These methods, based on threshold judgments at discrete time points, struggle to reflect the continuous dynamic evolution of physiological parameters like blood glucose and insulin. With the application of continuous monitoring technology, a large amount of patient physiological time-series data has been generated. Existing analytical methods typically extract statistical features using fixed-length windows or perform independent trend line fitting. Fixed windows cannot adapt to the inherent time-scale differences in various physiological processes. Separating statistical features from trend changes also prevents the formation of a unified trajectory view describing the continuous migration of health status, making traditional classification models insufficiently sensitive to identifying atypical disease courses.

[0003] In the application phase of subtyping models, conventional systems typically use pre-trained static models directly for inference on new data, with fixed feature extraction rules and model parameters. This process lacks a mechanism for self-adjustment based on individual real-time data. When faced with new cases exhibiting mixed features or at the subtype boundary, the system cannot optimize its representation of specific individual time-series data, potentially leading to poor stability and reduced confidence in the subtyping results. A method is needed to construct dynamic health trajectories from time-series data that integrate multi-scale information and trend changes, and to establish a mechanism for collaborative iterative optimization of trajectory modeling with preliminary subtyping judgment to improve the individualized judgment capability of diabetes subtyping. Summary of the Invention

[0004] The purpose of this invention is to provide a classification and diagnostic system for diabetes to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a diagnostic system for the subtyping of diabetes, the system comprising:

[0006] The data pre-analysis module imports the historical health records and real-time monitoring streams of diabetic patients and performs preprocessing to generate standardized patient time-series data.

[0007] The pattern trajectory module performs multi-scale sliding window segmentation on the standardized patient time-series data, calculates the statistical distribution characteristics and slope of the change trend of physiological indicators within each window, and constructs a pattern trajectory describing the individual's health evolution path.

[0008] The feature mapping module projects the pattern trajectory into a high-dimensional feature space, identifies the core nodes and boundary nodes of the trajectory cluster, and divides the discrete diabetes subtype prototypes.

[0009] The individual profiling module extracts newly generated data points from the real-time monitoring stream, calculates the multidimensional distance from the newly generated data points to each core node of the diabetes subtype prototype, evaluates the migration probability of the newly generated data points along each prototype trajectory direction, and generates a preliminary subtype membership vector.

[0010] The dynamic iteration module feeds back the preliminary subtype membership vector to the pattern trajectoryization module as the basis for adjusting the sliding window size and the weighting of trend calculation. After multiple iterations, it outputs a stable patient-specific health trajectory and an updated subtype membership vector.

[0011] The comprehensive analysis module integrates the stable patient-specific health trajectory with the updated subtype membership vector to generate a structured diagnostic opinion that includes specific subtype labels and confidence levels.

[0012] Preferably, the pattern tracing module includes:

[0013] For the standardized patient time-series data, an adaptive window width algorithm is applied to determine the boundaries of multiple sliding windows;

[0014] Within each defined sliding window, the empirical distribution functions of blood glucose and insulin concentration, along with their first and second moments, are calculated in parallel to obtain a window-level statistical summary.

[0015] Within each defined sliding window, the least squares method is used to fit linear segments of the changes in blood glucose and insulin concentration over time, and the slope and intercept of the linear segments are recorded as trend segments.

[0016] Window-level statistical summaries and trend segments of all windows are concatenated in chronological order to form a continuous pattern trajectory;

[0017] The continuous pattern trajectory is smoothed to eliminate abrupt changes at the window boundaries, resulting in the final pattern trajectory.

[0018] Preferably, the feature mapping module includes:

[0019] Using the window-level statistical summary and trend fragment of each window as the original features, the kernel function is used to transform them into vectors in a high-dimensional feature space;

[0020] Calculate the pairwise Euclidean distances between all vectors in the high-dimensional feature space and construct the distance matrix;

[0021] Based on the distance matrix, automatically find local density maxima and points that are extremely far from higher density points, and mark them as core nodes;

[0022] Around each core node, neighboring vectors are absorbed according to a preset distance threshold to form an initial cluster centered on the core node. Vectors that are not absorbed are marked as boundary nodes.

[0023] Initial clusters with overlap exceeding a threshold are merged, and the boundaries of each diabetes subtype prototype are defined based on the degree of ambiguity in the attribution of boundary nodes, thus completing the discrete diabetes subtype prototype classification.

[0024] Preferably, the individual profiling module includes:

[0025] Receive newly generated data points in the real-time monitoring stream and extract features similar to the window-level statistical summary;

[0026] Calculate the set of multidimensional Euclidean distances from the newly generated data point features to the core node features of each diabetes subtype prototype;

[0027] The relationship between the newly generated data points and their historical pattern trajectories is analyzed to simulate the possibility of their evolution along the prototype trajectory of each diabetes subtype. The migration probability is calculated using the transition probability matrix.

[0028] The multidimensional Euclidean distance set is normalized and weighted and fused with the corresponding migration probabilities to generate a preliminary subtype membership vector, where each dimension of the vector represents a probability score of belonging to a diabetes subtype prototype.

[0029] Preferably, the dynamic iteration module includes:

[0030] The preliminary subtype membership vector is input into the pattern trajectory module as a reference for adjusting the width of the next sliding window, and a more refined window division is adopted for the physiological index change range corresponding to the subtype prototype with high membership degree.

[0031] The standardized patient time-series data is re-segmented and recalculated using the adjusted sliding window to generate a new round of pattern trajectories;

[0032] The new pattern trajectory is input into the feature mapping module and the individual profile module again to recalculate the updated subtype membership vector;

[0033] Multiple iterative calculations are performed until the change norm of the subtype membership vector is less than a set threshold. Convergence is then determined, and the corresponding pattern trajectory is output as a stable patient-specific health trajectory. The subtype membership vector at this point is also output as the updated subtype membership vector.

[0034] Preferably, the comprehensive analysis module includes:

[0035] The clinical pathway knowledge graph contains association rules for typical pathological markers corresponding to different subtypes;

[0036] The stable patient-specific health trajectory is analyzed to identify abnormal inflection point sequences that exceed clinical thresholds.

[0037] The physiological indicators in the abnormal inflection point sequence are matched with the nodes in the clinical pathway knowledge graph, which stores the typical fluctuation range and order of appearance of various biomarkers under different diabetes subtypes in a graph structure.

[0038] Based on the successfully matched nodes and their connections in the graph, activate the corresponding diagnostic rule chain;

[0039] The conclusions of the activated diagnostic rule chain are cross-validated with the updated subtype membership vector. If the conclusions are consistent, the confidence level is increased; if there is a conflict, an arbitration mechanism based on rule confidence level is initiated.

[0040] Integrate the results of cross-validation or arbitration, and generate a formatted, structured diagnostic opinion that includes specific subtype labels and confidence levels.

[0041] Preferably, the data pre-analysis module includes:

[0042] The system receives historical health records and real-time monitoring streams from multiple sources, wherein the historical health records are from an electronic medical record database and the real-time monitoring streams are from a continuous glucose monitoring device.

[0043] The historical health records are timestamped and aligned with the real-time monitoring stream, and missing data points are interpolated.

[0044] The wavelet transform method is used to filter out high-frequency noise components in the aligned and interpolated data, while retaining low-frequency and trend components that conform to physiological changes.

[0045] The filtered data is standardized and scaled so that its mean and variance fall into a uniform dimension, thus obtaining the standardized patient time series data.

[0046] Preferably, the step of applying an adaptive window width algorithm to the standardized patient time-series data to determine the boundaries of multiple sliding windows includes:

[0047] Calculate the local gradient change rate for each data point in the normalized patient time series data, and generate a window width adjustment coefficient based on the local gradient change rate;

[0048] The starting and ending points of the sliding window are dynamically divided according to the window width adjustment coefficient. The region with a high gradient rate of change corresponds to a smaller window width to capture detailed features, while the region with a low gradient rate of change corresponds to a larger window width to smooth noise.

[0049] A sliding window boundary optimization algorithm is used to eliminate overlaps or gaps between windows, ensuring that the window sequence continuously covers the entire time series data range.

[0050] Verify the statistical consistency of data points within each sliding window. If there is a discrepancy, readjust the window boundaries until the preset consistency threshold is met.

[0051] Preferably, the step of mapping the window-level statistical summary and trend fragment of each window to a vector in a high-dimensional feature space through kernel function transformation includes:

[0052] Extract the window-level statistical summary and trend fragment of each sliding window as the original feature vector, which includes mean, variance and slope parameters;

[0053] The radial basis function kernel is selected as the mapping function to calculate the similarity measure between the original feature vector and the preset feature anchor point;

[0054] Based on the similarity metric, the original feature vector is nonlinearly transformed to a high-dimensional feature space to obtain a high-dimensional feature vector;

[0055] The high-dimensional feature vectors are normalized to eliminate dimensional differences and preserve the relative distance relationships between features.

[0056] Preferably, the step of using wavelet transform to filter out high-frequency noise components in the aligned and interpolated data while retaining low-frequency and trend components that conform to physiological changes includes:

[0057] Select wavelet basis functions suitable for physiological signals, perform multi-resolution wavelet decomposition on the aligned and interpolated time series data, and obtain wavelet coefficients of different frequency sub-bands;

[0058] Set a noise threshold, and set the coefficients in the high-frequency subband whose absolute value of wavelet coefficients is lower than the noise threshold to zero to filter out high-frequency noise;

[0059] Perform inverse wavelet transform on the wavelet coefficients of the low-frequency subband to reconstruct the noise-filtered time series data;

[0060] Verify the consistency of trends between the reconstructed data and the original data to ensure that the low-frequency components of physiological change patterns are fully preserved.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] By segmenting standardized patient time-series data using a multi-scale sliding window and simultaneously calculating the statistical distribution characteristics and trend slope of physiological indicators within each window, a pattern trajectory integrating static features and dynamic change rates is constructed. This trajectory characterizes health status at multiple time resolutions, simultaneously capturing the details of short-term fluctuations and the long-term evolution direction of indicators. Compared to using only fixed windows to extract features or performing trend analysis alone, the generated health path representation is more continuous and richer in information dimensions. This enables the system to more sensitively identify key states such as prediabetes and blood glucose fluctuation patterns, distinguishing patients at different stages of progression and fluctuation types, providing a higher-quality and more discriminative dynamic profile foundation for subsequent subtyping.

[0063] The preliminary subtype membership vector generated by the individual profiling module is fed back to the pattern trajectoryization module as the basis for adjusting the sliding window size and trend calculation weighting, forming a closed loop from initial judgment to feature extraction parameter optimization. The preliminary membership reflects the initial correlation strength between new data points and each subtype prototype. Adjusting the window size based on this essentially aligns the time scale of feature extraction with the typical evolution rhythm of the individual's most likely subtype. Introducing weighting into trend calculation highlights key change stages more relevant to diagnosis in the time dimension, based on the preliminary subtyping tendency. After multiple iterations, the system's output patient-specific health trajectory and final membership vector reach a stable state. This achieves synergistic optimization of feature extraction and classification judgment, avoiding the risk of misjudging atypical individuals by fixed-parameter models. Through iterative self-adjustment, the system makes the generated trajectory more reflective of the individual's unique evolutionary logic, thereby improving the personalization of subtyping and the reliability of the final diagnostic opinion, and enhancing the ability to identify cases with mixed features or in the subtype transition period. Attached Figure Description

[0064] Figure 1 This is a schematic diagram illustrating the working principle of the diabetes classification and diagnostic system described in this invention.

[0065] Figure 2 A flowchart for the pattern tracing module;

[0066] Figure 3 This is a flowchart of the feature mapping module;

[0067] Figure 4 A bar chart showing the correlation between diabetes subtypes;

[0068] Figure 5 Multi-series time trend graphs of physiological indicators for diabetic patients. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0070] Please see Figure 1 This invention provides a diabetes subtype diagnostic system, comprising: a data pre-analysis module responsible for importing and standardizing historical health records from an electronic medical record database and real-time monitoring streams from continuous glucose monitoring devices, generating standardized patient time-series data; a pattern trajectoryization module receiving the standardized time-series data, segmenting it using multi-scale sliding window technology, and calculating the statistical distribution characteristics and trend slope of key physiological indicators such as blood glucose and insulin within each window, and then concatenating these window features to construct a continuous pattern trajectory describing the evolution path of an individual's health status; a feature mapping module projecting the constructed pattern trajectory onto a high-dimensional feature space using methods such as kernel functions, identifying core nodes and boundary nodes in the trajectory points within this space, and dividing multiple discrete diabetes subtype prototypes based on the distance and density relationships between nodes; and an individual profiling module extracting features from newly generated data points in the real-time monitoring stream and calculating their multidimensional distances to the core nodes of each subtype prototype, while simultaneously assessing the probability of the data point migrating along the predetermined trajectory direction of each prototype, fusing distance and probability information to generate a preliminary subtype membership vector. The dynamic iteration module feeds back the initial subtype membership vector to the pattern trajectoryization module, serving as the basis for adjusting the subsequent sliding window size and trend calculation weights. After multiple rounds of iterative calculations until the membership vector stabilizes, it finally outputs a stable patient-specific health trajectory and an updated subtype membership vector. The comprehensive judgment module integrates the above stable health trajectory and updated membership vector, and combines it with the clinical pathway knowledge graph for logical reasoning and cross-validation, ultimately generating a structured diagnostic opinion containing specific subtype labels and their confidence levels.

[0071] Example 1: See Figure 2In the implementation of the pattern tracing module, an adaptive window width algorithm is applied to the standardized patient time-series data to determine the boundaries of multiple sliding windows. This algorithm calculates the local gradient rate of change for each data point in the time-series data and generates a window width adjustment coefficient based on this. The starting and ending points of the sliding windows are dynamically divided according to the window width adjustment coefficient. Smaller window widths are used in regions with high gradient rates of change to capture detailed features, while larger window widths are used in regions with low gradient rates of change to smooth noise. A sliding window boundary optimization algorithm is used to eliminate overlaps or gaps between windows, ensuring that the window sequence continuously covers the entire time-series data range. The statistical consistency of data points within each sliding window is verified; if inconsistencies are found, the window boundaries are readjusted until a preset consistency threshold is met. Within each determined sliding window, the empirical distribution functions of blood glucose and insulin concentration indicators, along with their first and second moments, are calculated in parallel to obtain a window-level statistical summary. Within each determined sliding window, the least squares method is used to fit linear segments of blood glucose and insulin concentration indicators changing over time, and the slope and intercept of these linear segments are recorded as trend segments. Window-level statistical summaries and trend segments from all windows are concatenated in chronological order to form a continuous pattern trajectory. The continuous pattern trajectory is then smoothed to eliminate abrupt changes at window boundaries, yielding the final pattern trajectory.

[0072] In practice, the pattern tracing module receives normalized patient time-series data from the data pre-analysis module and begins multi-scale sliding window segmentation and pattern trajectory construction. Taking a simulated blood glucose value sequence containing 72 consecutive hours of data with 5-minute sampling intervals as an example, this sequence is a crucial component of the normalized patient time-series data. The pattern tracing module first applies an adaptive window width algorithm to this normalized patient time-series data to determine the boundaries of multiple sliding windows. The specific process of the adaptive window width algorithm is to calculate the local gradient change rate of each data point in the normalized patient time-series data and generate a window width adjustment coefficient based on the local gradient change rate of each data point in the normalized patient time-series data. In some embodiments, the calculation of the window width adjustment coefficient can be based on the following relationship:

[0073] ;

[0074] in: This represents the suggested window width adjustment factor for the region near the i-th data point. This represents the rate of change of the local gradient at the i-th data point. It is a constant factor related to the width of the baseline window. Local gradient rate of change. The adjustment coefficient for the window width is obtained by calculating the sum of the squares of the differences between a data point and its immediate neighbors. Regions with high gradient rate of change correspond to smaller window width adjustment coefficients, while regions with low gradient rate of change correspond to larger window width adjustment coefficients.

[0075] The sliding window dynamically divides its start and end points based on a window width adjustment coefficient. Smaller window widths are used in regions with high gradient change rates to capture detailed features, while larger window widths are used in regions with low gradient change rates to smooth noise. The actual window width is determined by the product of the baseline width and the window width adjustment coefficient. Therefore, the rapid rise and fall in blood glucose levels after a meal will be divided into a series of narrower windows, while the relatively stable nighttime phase will be divided into wider windows. A sliding window boundary optimization algorithm is used to eliminate overlaps or gaps between windows, ensuring that the window sequence continuously covers the entire time-series data range. This algorithm compares the timestamps of the end and start points of adjacent windows and fine-tunes them to ensure that the end point of one window and the start point of the next window are precisely aligned in time. The statistical consistency of data points within each sliding window is verified; if inconsistencies are found, the window boundaries are readjusted until a preset consistency threshold is met. Statistical consistency verification is achieved by checking whether the variance of data points within a window exceeds a threshold calculated based on the overall sequence variance. For windows that exceed the threshold, the adaptive window width algorithm will split them into smaller windows until the data fluctuations within each sub-window meet the consistency requirements.

[0076] In practice, within each sliding window whose boundaries are determined by an adaptive window width algorithm, the empirical distribution functions of blood glucose and insulin concentration, along with their first and second moments, are calculated in parallel to obtain a window-level statistical summary. This summary includes statistics such as the mean, variance, and skewness of blood glucose within the window's time range. Within each defined sliding window, a linear segment of the changes in blood glucose and insulin concentration over time is fitted using the least squares method. The slope and intercept of this linear segment are recorded as trend segments. The slope in the trend segment quantifies whether the physiological indicators are rising, falling, or remaining stable within that window, and the rate of change. The window-level statistical summaries and trend segments of all windows are concatenated chronologically to form a continuous pattern trajectory. This continuous pattern trajectory, in the time dimension, is a sequence of alternating statistical and trend feature vectors, characterizing the evolution of the patient's physiological state over time. Smoothing is then applied to the continuous pattern trajectory to eliminate abrupt changes at window boundaries, resulting in the final pattern trajectory. In some embodiments, the smoothing process employs a moving average filter to filter the statistical summary features of adjacent windows, resulting in a smoother feature transition from one window to the next, avoiding feature value jumps caused by window boundary division. Optionally, the weights of the smoothing process can be adjusted according to the window width, with wider windows being assigned higher weights during smoothing. It is understood that the final pattern trajectory after smoothing can more coherently represent the evolution of the patient's health status, providing stable input for subsequent feature mapping and subtype classification.

[0077] Example 2: See Figure 3 In the implementation of the feature mapping module, the window-level statistical summary and trend fragment of each window are used as the original features, and they are mapped to vectors in a high-dimensional feature space through kernel function transformation. The window-level statistical summary and trend fragment of each sliding window are extracted as the original feature vector, which includes mean, variance, and slope parameters. A radial basis function kernel is selected as the mapping function to calculate the similarity metric between the original feature vector and the preset feature anchor points. Based on this similarity metric, the original feature vector is nonlinearly transformed to a high-dimensional feature space to obtain a high-dimensional feature vector. The high-dimensional feature vector is normalized to eliminate dimensional differences and retain the relative distance relationships between features. The Euclidean distance between all pairs of vectors in the high-dimensional feature space is calculated to construct a distance matrix. Based on this distance matrix, points with local density maxima and points with maximum distances to higher density points are automatically found and marked as core nodes. Around each core node, neighboring vectors are absorbed according to a preset distance threshold to form an initial cluster centered on the core node; vectors not absorbed are marked as boundary nodes. Initial clusters with overlap exceeding a threshold are merged, and the boundaries of each diabetes subtype prototype are defined based on the degree of ambiguity in the attribution of boundary nodes, thus completing the discrete diabetes subtype prototype classification.

[0078] In specific implementation, the feature mapping module receives the final pattern trajectory from the pattern trajectoryization module. The pattern trajectory is formed by concatenating a series of window-level statistical summaries and trend segments in chronological order. Taking a simulated dataset containing pattern trajectories of one hundred diabetic patients as an example, each patient's pattern trajectory contains dozens of windows. The window-level statistical summary and trend segment of each window together constitute an original feature vector. The feature mapping module first extracts the window-level statistical summary and trend segment of each sliding window as the original feature vector. The original feature vector contains mean, variance, and slope parameters. The radial basis function kernel function is selected as the mapping function. The similarity measure between the original feature vector and the preset feature anchor point is calculated. In some embodiments, the preset feature anchor point is a representative feature vector randomly selected from the training dataset or determined by clustering methods. Based on the similarity measure, the original feature vector is nonlinearly transformed to a high-dimensional feature space to obtain a high-dimensional feature vector. The high-dimensional feature vector is normalized to eliminate dimensional differences and retain the relative distance relationship between features.

[0079] In practical implementation, the Euclidean distance between all pairwise high-dimensional feature vectors in the high-dimensional feature space is calculated to construct a distance matrix. This distance matrix is ​​a symmetric matrix whose elements represent the degree of separation between any two high-dimensional feature vectors in the feature space. Based on the distance matrix, points with local density maxima and points extremely far from higher-density points are automatically identified and marked as core nodes. It can be understood that local density maxima represent the centers of densely distributed regions in the feature space, while points extremely far from higher-density points may be separating points between different dense regions. The process of finding core nodes involves calculating the local density value of each high-dimensional feature vector and the minimum distance from a high-dimensional feature vector to a higher-density feature vector. In some embodiments, the local density value can be calculated using the following formula:

[0080] ;

[0081] in: This represents the local density value of the m-th high-dimensional eigenvector. This represents the Euclidean distance between the m-th and n-th high-dimensional eigenvectors in the distance matrix. It is a bandwidth parameter used to control the density decay rate, high local density value. This indicates that the m-th high-dimensional eigenvector is surrounded by a large number of other high-dimensional eigenvectors. For each high-dimensional eigenvector, find all high-dimensional eigenvectors with higher local density values ​​and calculate the minimum distance to these high-dimensional eigenvectors, denoted as . Local density value Extremely high and at a distance The extremely large high-dimensional feature vectors were also selected as core nodes.

[0082] In practice, around each core node, neighboring high-dimensional feature vectors are absorbed according to a preset distance threshold to form an initial cluster centered on the core node. The preset distance threshold can be a fixed value set based on the statistical characteristics of the distance matrix, or a dynamic value adaptively adjusted according to the local density of the core node. High-dimensional feature vectors not absorbed by any core node are marked as boundary nodes. Boundary nodes may be located in the intersection area of ​​different initial clusters, and their classification is ambiguous. Initial clusters with an overlap exceeding the threshold are merged. The overlap is measured by calculating the ratio of the number of shared high-dimensional feature vectors between two initial clusters to the size of the smaller initial cluster. The boundaries of each diabetes subtype prototype are delineated according to the degree of ambiguity in the classification of the boundary nodes, thus completing the discrete classification of diabetes subtype prototypes. It can be understood that the boundary node region with a high degree of classification ambiguity corresponds to the fuzzy transition zone between diabetes subtype prototypes, while the region with a clear classification defines the core range of the diabetes subtype prototype. Optionally, the final classification of the boundary nodes can be determined by calculating the weighted distance from the boundary nodes to each core node. The weight is related to the tightness of the initial cluster to which the core node is located. Optionally, the number of diabetes subtype prototypes can be automatically determined during the core node identification stage without the need for pre-setting.

[0083] Example 3: In the implementation of the individual profiling module, newly generated data points are received from the real-time monitoring stream, and features similar to those in the window-level statistical summary are extracted. The multidimensional Euclidean distance set from the features of the newly generated data points to the features of the core nodes of each diabetes subtype prototype is calculated. The relationship between the newly generated data points and their historical patterns is analyzed, simulating the possibility of their evolution along the trajectory of each diabetes subtype prototype. The migration probability is calculated using a transition probability matrix. The multidimensional Euclidean distance set is normalized and weighted and fused with the corresponding migration probabilities to generate a preliminary subtype membership vector, where each dimension of the vector represents a probability score of belonging to a diabetes subtype prototype.

[0084] In the implementation of the dynamic iteration module, the initial subtype membership vector is input into the pattern trajectoryization module as a reference for adjusting the width of the sliding window for the next iteration. A more refined window division is used for the physiological indicator variation range corresponding to subtype prototypes with high membership degrees. The adjusted sliding window is then used to re-segment and recalculate the standardized patient time-series data, generating a new round of pattern trajectories. These new pattern trajectories are then input again into the feature mapping module and the individual profiling module to recalculate the updated subtype membership vector. This iterative calculation is performed multiple times until the norm of the subtype membership vector is less than a set threshold, at which point convergence is considered achieved. The corresponding pattern trajectory at this point is output as a stable patient-specific health trajectory, and the subtype membership vector at this point is output as the updated subtype membership vector.

[0085] In practice, the individual profiling module receives newly generated data points from the real-time monitoring stream. These newly generated data points can be, for example, blood glucose and insulin concentration values ​​acquired in the most recent monitoring period. For a diabetic patient with a final pattern trajectory constructed from historical data, the individual profiling module extracts features similar to the window-level statistical summary within the time window corresponding to the newly generated data points. These features include the mean, variance, and slope of the blood glucose value. The module calculates the multidimensional Euclidean distance set from the features of the newly generated data points to the core node features of each diabetic subtype prototype. The core node features of the diabetic subtype prototype come from the results of the feature mapping module. Each diabetic subtype prototype typically has one or more core node feature vectors. The multidimensional Euclidean distance set contains the distance values ​​from the features of the newly generated data points to all core nodes of all diabetic subtype prototypes. The relationship between newly generated data points and historical pattern trajectories is analyzed to simulate the probability of these data points evolving along the trajectory directions of each diabetes subtype prototype. The migration probability is calculated using a transition probability matrix, which defines the likelihood of transitioning from the current state of the historical pattern trajectory to the typical state of each diabetes subtype prototype. This matrix can be obtained through statistical learning based on the trajectory evolution patterns of different subtype patients in historical data. The multidimensional Euclidean distance set is normalized and weighted with the corresponding migration probabilities to generate a preliminary subtype membership vector, where each dimension of the vector represents a probability score of belonging to a diabetes subtype prototype. In some embodiments, the score of the k-th dimension of the preliminary subtype membership vector... It can be calculated in the following ways:

[0086] ;

[0087] in: This represents the initial score indicating that the newly generated data point belongs to the k-th diabetes subtype prototype. This represents the normalized value of the minimum Euclidean distance from the newly generated data point features to all core node features of the k-th diabetes subtype prototype. This represents the probability that a newly generated data point will migrate to the prototype of the k-th diabetes subtype. This is a fusion weight coefficient between 0 and 1, used to balance the contributions of distance information and trend migration information in the scoring. It can be understood that the normalization operation maps distance to the [0,1] interval; the closer the distance, the higher the weight. The closer the value is to 1, the higher the migration probability. The higher the score, the higher the score. The newly generated data points obtained through weighted fusion form the preliminary score vector of each diabetes subtype prototype, i.e., the preliminary subtype membership vector.

[0088] In practice, the dynamic iteration module inputs the initial subtype membership vector into the pattern tracing module as a reference for adjusting the width of the sliding window for the next iteration. For subtype prototypes with high membership, a more refined window division is used for the physiological indicator variation range. For example, if the initial subtype membership vector shows that the patient is highly inclined towards a certain diabetes subtype prototype characterized by dramatic postprandial blood glucose fluctuations, the pattern tracing module will use a smaller window width within the identified postprandial time period to more precisely analyze the details of rapid blood glucose changes. The adjusted sliding window is then used to re-segment and calculate the standardized patient time-series data, generating a new round of pattern trajectories. Due to the adjustment of the window segmentation strategy, the window-level statistical summary and the level of detail in the trend segments of the new round of pattern trajectories may change. The new round of pattern trajectories is then input again into the feature mapping module and the individual profiling module to recalculate the updated subtype membership vector. This process optimizes the patient feature representation and subtype assessment based on initial diagnostic feedback. Multiple iterative calculations are performed until the norm of change of the subtype membership vector is less than a set threshold, at which point convergence is considered achieved. In some embodiments, the norm of change is calculated as the sum of squares of the differences in each dimension between the current round's subtype membership vector and the previous round's subtype membership vector. When this sum of squares is less than a preset minimum positive number, the subtype membership vector is considered stable. The corresponding pattern trajectory at this point is output as the stable patient-specific health trajectory, and the current subtype membership vector is output as the updated subtype membership vector. It can be understood that the dynamic iteration module, through feedback loops, enables the system to adapt to the characteristics of individual data. Initially, the window segmentation may be inaccurate, but iterative processes optimize this process, resulting in extracted pattern trajectories that better reflect the patient's true physiological patterns, thereby improving the accuracy of subtype identification. Optionally, the threshold can be calibrated based on historical data convergence, and optionally, an upper limit can be set for the number of iterations to prevent infinite loops.

[0089] See Figure 4 This is a bar chart comparing diabetes subtypes. The dual-series bar chart compares two metrics: "normalized Euclidean distance" and "transfer probability" for the four diabetes subtypes. Type 2 diabetes has the highest "normalized Euclidean distance" (close to 0.9), indicating the highest match between the new data point and the core features of this subtype. The "transfer probability" for each subtype is generally lower than the corresponding "normalized Euclidean distance," with type 2 diabetes showing a relatively higher transfer probability. This chart is typically used in the individual profiling module of diabetes classification diagnosis to help determine the diabetes subtype a new data point belongs to.

[0090] Example 4: In the implementation of the comprehensive judgment module, the clinical pathway knowledge graph contains association rules for typical pathological markers corresponding to different subtypes. Stable patient-specific health trajectories are analyzed, and abnormal inflection point sequences exceeding clinical thresholds are identified. Physiological indicators in these abnormal inflection point sequences are matched with nodes in the clinical pathway knowledge graph, which stores the typical fluctuation ranges and order of appearance of various markers under different diabetes subtypes in a graph structure. Based on successfully matched nodes and their connections in the graph, the corresponding diagnostic rule chains are activated. The conclusions of the activated diagnostic rule chains are cross-validated with the updated subtype membership vector. If the conclusions are consistent, the confidence level is increased; if there is a conflict, an arbitration mechanism based on rule confidence level is initiated. The results of cross-validation or arbitration are integrated to generate a structured diagnostic opinion containing specific subtype labels and confidence levels.

[0091] In practical implementation, the comprehensive judgment module receives stable patient-specific health trajectories and updated subtype membership vectors from the dynamic iteration module. The stable patient-specific health trajectory is a series of physiological state feature points arranged chronologically. The updated subtype membership vector is a list of values, where each value represents a probability score indicating that the patient belongs to a specific diabetes subtype. The clinical pathway knowledge graph of the comprehensive judgment module contains association rules for typical pathological biomarkers corresponding to different subtypes. The clinical pathway knowledge graph stores the typical fluctuation ranges and order of appearance of various biomarkers under different diabetes subtypes in a graph structure. Nodes in the graph represent physiological states or clinical events, and edges represent transition relationships or logical connections between states. Refer to Table 1 for a simplified clinical pathway knowledge graph, which includes some rules related to the two hypothetical subtypes.

[0092] Table 1: Nodes and Association Rules of Clinical Pathway Knowledge Graph

[0093]

[0094] In practice, a stable patient-specific health trajectory is analyzed, and abnormal inflection point sequences exceeding clinical thresholds are identified. These abnormal inflection point sequences are points or consecutive points in the trajectory where the statistical characteristics or trend slope of physiological indicators change drastically and are marked as abnormal after comparison with preset clinical thresholds. The physiological indicators in the abnormal inflection point sequences are matched with nodes in the clinical pathway knowledge graph. This matching process involves comparing the physiological indicator type and its quantified value at each inflection point in the abnormal inflection point sequence with the "physiological indicators" and "abnormal conditions" of each node in the clinical pathway knowledge graph. Based on the successfully matched nodes and their connections in the clinical pathway knowledge graph, the corresponding diagnostic rule chain is activated. For example, if node N1 is matched, and subsequently node N2 is matched in time sequence, a diagnostic rule chain pointing to type A diabetes is activated. In some embodiments, the activated diagnostic rule chain outputs a subtype inference result based on graph logic and its graph confidence level.

[0095] In practice, the conclusions of the activated diagnostic rule chain are cross-validated with the updated subtype membership vector. If the conclusions are consistent, the confidence level is increased; if a conflict exists, an arbitration mechanism based on rule confidence is initiated. Cross-validation compares the subtype inference result output by the diagnostic rule chain with the subtype with the highest probability score in the updated subtype membership vector. Increasing the confidence level can be achieved through weighted averaging. It can be understood that the conclusions of the diagnostic rule chain originate from a graphical representation of clinical experience and pathological logic, while the updated subtype membership vector originates from data-driven trajectory and prototype matching. Consistency between the two enhances the reliability of the final diagnosis. In some embodiments, when a conflict exists, the arbitration mechanism adjudicates based on the graphical confidence level of the diagnostic rule chain and the scores of the corresponding dimensions in the updated subtype membership vector. The final confidence score after arbitration can be calculated using the following formula. :

[0096] ;

[0097] in: This represents the final confidence score assigned to a subtype label after arbitration. This represents the confidence level of the diagnostic rule chain output in the clinical pathway knowledge graph for that subtype. This represents the probability score corresponding to the subtype in the updated subtype membership vector. This is an arbitration weighting coefficient between 0 and 1, used to adjust the relative importance of the graph logic and data-driven results in the final decision. The results of cross-validation or arbitration are integrated and formatted to generate a structured diagnostic opinion containing specific subtype labels and confidence scores. This structured diagnostic opinion can be understood as a machine-readable and human-interpretable data object, such as a JSON string that explicitly lists the most likely diabetes subtype, the corresponding confidence score, and other candidate subtype information.

[0098] Example 5: In the implementation of the data pre-analysis module, historical health records and real-time monitoring streams from multiple sources are received. The historical health records come from an electronic medical record database, and the real-time monitoring streams come from continuous glucose monitoring devices. The historical health records and real-time monitoring streams are timestamped, and missing data points are imputed. Wavelet transform is used to filter out high-frequency noise components in the aligned and imputed data, retaining low-frequency and trend components consistent with physiological changes. A suitable wavelet basis function for physiological signals is selected, and multi-resolution wavelet decomposition is performed on the aligned and imputed time-series data to obtain wavelet coefficients for different frequency sub-bands. A noise threshold is set, and coefficients in the high-frequency sub-bands whose absolute values ​​are below the noise threshold are set to zero to filter out high-frequency noise. Inverse wavelet transform is performed on the wavelet coefficients of the low-frequency sub-bands to reconstruct the noise-filtered time-series data. The trend consistency between the reconstructed data and the original data is verified to ensure that the low-frequency components of physiological changes are completely preserved. The noise-filtered data is standardized and scaled so that its mean and variance fall into a uniform dimension, resulting in standardized patient time-series data.

[0099] In practice, the data pre-analysis module receives historical health records and real-time monitoring streams from multiple sources. The historical health records come from an electronic medical record database, while the real-time monitoring streams come from a continuous glucose monitoring device. For example, the electronic medical record database may provide discrete test records such as glycated hemoglobin and fasting blood glucose over the past year, while the continuous glucose monitoring device provides continuous readings of subcutaneous interstitial fluid glucose concentration at a frequency of one minute. The module timestamps the historical health records and the real-time monitoring streams and interpolates missing data points. Timestamp alignment unifies data from different sources onto a common time reference. Interpolation of missing data points can use linear interpolation or prediction methods based on historical patterns to fill in short-term data gaps caused by device offline or transmission failure.

[0100] High-frequency noise components in the aligned and interpolated data are filtered out using wavelet transform, while low-frequency and trend components consistent with physiological changes are retained. Wavelet basis functions suitable for physiological signals are selected, such as Daubechies or Symlets wavelets. Multi-resolution wavelet decomposition is then performed on the aligned and interpolated time-series data to obtain wavelet coefficients for different frequency sub-bands. In some embodiments, the number of decomposition levels L can be determined based on the main frequency components of the signal, expressed by the formula:

[0101] ;

[0102] in: This indicates the number of wavelet decomposition levels. This indicates the length of the aligned and interpolated timing data. It is an empirical constant used to prevent over-decomposition. This represents the floor function. A noise threshold is set, and coefficients in the high-frequency sub-band whose absolute values ​​are below the noise threshold are set to zero to filter out high-frequency noise. The noise threshold can be adaptively determined based on the statistical distribution of the wavelet coefficients in the highest-frequency sub-band. An inverse wavelet transform is performed on the wavelet coefficients of the low-frequency sub-band to reconstruct the denoised time-series data. The reconstruction process utilizes the wavelet coefficients of all high-frequency and low-frequency sub-bands, and recovers the time-domain signal using a wavelet synthesis algorithm. The trend consistency between the reconstructed data and the original data is verified to ensure that the low-frequency components of the physiological change patterns are fully preserved. Trend consistency verification can be achieved by comparing whether the slopes of the linear fitting between the reconstructed data and the original data remain consistent within the sliding window.

[0103] In practice, the filtered data undergoes standardization and scaling to ensure that its mean and variance fall within a unified dimension, resulting in standardized patient time-series data. Standardization and scaling can employ the Z-score standardization method, subtracting the mean from each physiological indicator sequence and then dividing by the standard deviation, resulting in a mean of 0 and a variance of 1 for the processed data. In essence, after timestamp alignment, missing value imputation, wavelet transform filtering, and standardization and scaling, historical health records from the electronic medical record database and real-time monitoring streams from continuous glucose monitoring devices are integrated and transformed into a set of standardized patient time-series data with a unified time reference, continuous data, and consistent dimensions. This standardized patient time-series data provides high-quality input for the subsequent pattern tracing module.

[0104] See Figure 5 This is a multi-series time-trend graph of physiological indicators in diabetic patients. Fasting blood glucose (blue) and insulin levels (green) fluctuate highly synchronously, exhibiting a cyclical pattern of alternating peaks and troughs. Glycated hemoglobin (red) remains generally stable (around 5%), showing no significant changes with short-term fluctuations in blood glucose / insulin levels. Fasting blood glucose repeatedly approaches a high value of 160 mg / dL, with corresponding insulin levels peaking at around 20 μU / mL, consistent with the physiological response of diabetic patients: "elevated blood glucose → compensatory insulin secretion." This graph is commonly used for time-series data monitoring in diabetes management to assist in analyzing the dynamic correlation between blood glucose and insulin levels and the stability of the patient's condition.

[0105] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0106] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A diagnostic system for the subtyping of diabetes, characterized in that, The system includes: The data pre-analysis module imports the historical health records and real-time monitoring streams of diabetic patients and performs preprocessing to generate standardized patient time-series data. The pattern trajectory module performs multi-scale sliding window segmentation on the standardized patient time-series data, calculates the statistical distribution characteristics and slope of the change trend of physiological indicators within each window, and constructs a pattern trajectory describing the individual's health evolution path. The feature mapping module projects the pattern trajectory into a high-dimensional feature space, identifies the core nodes and boundary nodes of the trajectory cluster, and divides the discrete diabetes subtype prototypes. The individual profiling module extracts newly generated data points from the real-time monitoring stream, calculates the multidimensional distance from the newly generated data points to each core node of the diabetes subtype prototype, evaluates the migration probability of the newly generated data points along each prototype trajectory direction, and generates a preliminary subtype membership vector. The dynamic iteration module feeds back the preliminary subtype membership vector to the pattern trajectoryization module as the basis for adjusting the sliding window size and the weighting of trend calculation. After multiple iterations, it outputs a stable patient-specific health trajectory and an updated subtype membership vector. The comprehensive analysis module integrates the stable patient-specific health trajectory with the updated subtype membership vector to generate a structured diagnostic opinion that includes specific subtype labels and confidence levels.

2. The diabetes classification and diagnostic system according to claim 1, characterized in that, The pattern tracing module includes: For the standardized patient time-series data, an adaptive window width algorithm is applied to determine the boundaries of multiple sliding windows; Within each defined sliding window, the empirical distribution functions of blood glucose and insulin concentration, along with their first and second moments, are calculated in parallel to obtain a window-level statistical summary. Within each defined sliding window, the least squares method is used to fit linear segments of the changes in blood glucose and insulin concentration over time, and the slope and intercept of the linear segments are recorded as trend segments. Window-level statistical summaries and trend segments of all windows are concatenated in chronological order to form a continuous pattern trajectory; The continuous pattern trajectory is smoothed to eliminate abrupt changes at the window boundaries, resulting in the final pattern trajectory.

3. The diabetes classification and diagnostic system according to claim 2, characterized in that, The feature mapping module includes: Using the window-level statistical summary and trend fragment of each window as the original features, the kernel function is used to transform them into vectors in a high-dimensional feature space; Calculate the pairwise Euclidean distances between all vectors in the high-dimensional feature space and construct the distance matrix; Based on the distance matrix, automatically find local density maxima and points that are extremely far from higher density points, and mark them as core nodes; Around each core node, neighboring vectors are absorbed according to a preset distance threshold to form an initial cluster centered on the core node. Vectors that are not absorbed are marked as boundary nodes. Initial clusters with overlap exceeding a threshold are merged, and the boundaries of each diabetes subtype prototype are defined based on the degree of ambiguity in the attribution of boundary nodes, thus completing the discrete diabetes subtype prototype classification.

4. The diabetes classification and diagnostic system according to claim 3, characterized in that, The individual profile module includes: Receive newly generated data points in the real-time monitoring stream and extract features similar to the window-level statistical summary; Calculate the set of multidimensional Euclidean distances from the newly generated data point features to the core node features of each diabetes subtype prototype; The relationship between the newly generated data points and their historical pattern trajectories is analyzed to simulate the possibility of their evolution along the prototype trajectory of each diabetes subtype. The migration probability is calculated using the transition probability matrix. The multidimensional Euclidean distance set is normalized and weighted and fused with the corresponding migration probabilities to generate a preliminary subtype membership vector, where each dimension of the vector represents a probability score of belonging to a diabetes subtype prototype.

5. The diabetes classification and diagnostic system according to claim 4, characterized in that, The dynamic iteration module includes: The preliminary subtype membership vector is input into the pattern trajectory module as a reference for adjusting the width of the next sliding window, and a more refined window division is adopted for the physiological index change range corresponding to the subtype prototype with high membership degree. The standardized patient time-series data is re-segmented and recalculated using the adjusted sliding window to generate a new round of pattern trajectories; The new pattern trajectory is input into the feature mapping module and the individual profile module again to recalculate the updated subtype membership vector; Multiple iterative calculations are performed until the change norm of the subtype membership vector is less than a set threshold. Convergence is then determined, and the corresponding pattern trajectory is output as a stable patient-specific health trajectory. The subtype membership vector at this point is also output as the updated subtype membership vector.

6. The diabetes classification and diagnostic system according to claim 5, characterized in that, The comprehensive analysis module includes: The clinical pathway knowledge graph contains association rules for typical pathological markers corresponding to different subtypes; The stable patient-specific health trajectory is analyzed to identify abnormal inflection point sequences that exceed clinical thresholds. The physiological indicators in the abnormal inflection point sequence are matched with the nodes in the clinical pathway knowledge graph, which stores the typical fluctuation range and order of appearance of various biomarkers under different diabetes subtypes in a graph structure. Based on the successfully matched nodes and their connections in the clinical pathway knowledge graph, activate the corresponding diagnostic rule chain; The conclusions of the activated diagnostic rule chain are cross-validated with the updated subtype membership vector. If the conclusions are consistent, the confidence level is increased; if there is a conflict, an arbitration mechanism based on rule confidence level is initiated. Integrate the results of cross-validation or arbitration, and generate a formatted, structured diagnostic opinion that includes specific subtype labels and confidence levels.

7. The diabetes classification and diagnostic system according to claim 1, characterized in that, The data pre-analysis module includes: The system receives historical health records and real-time monitoring streams from multiple sources, wherein the historical health records are from an electronic medical record database and the real-time monitoring streams are from a continuous glucose monitoring device. The historical health records are timestamped and aligned with the real-time monitoring stream, and missing data points are interpolated. The wavelet transform method is used to filter out high-frequency noise components in the aligned and interpolated data, while retaining low-frequency and trend components that conform to physiological changes. The filtered data is standardized and scaled so that its mean and variance fall into a uniform dimension, thus obtaining the standardized patient time series data.

8. The diabetes classification and diagnostic system according to claim 2, characterized in that, The step of applying an adaptive window width algorithm to the standardized patient time-series data to determine the boundaries of multiple sliding windows includes: Calculate the local gradient change rate for each data point in the normalized patient time series data, and generate a window width adjustment coefficient based on the local gradient change rate; The starting and ending points of the sliding window are dynamically divided according to the window width adjustment coefficient. The region with a high gradient rate of change corresponds to a smaller window width to capture detailed features, while the region with a low gradient rate of change corresponds to a larger window width to smooth noise. A sliding window boundary optimization algorithm is used to eliminate overlaps or gaps between windows, ensuring that the window sequence continuously covers the entire time series data range. Verify the statistical consistency of data points within each sliding window. If there is a discrepancy, readjust the window boundaries until the preset consistency threshold is met.

9. The diabetes classification and diagnostic system according to claim 3, characterized in that, The process of mapping the window-level statistical summary and trend fragments of each window to a vector in a high-dimensional feature space through kernel function transformation includes: Extract the window-level statistical summary and trend fragment of each sliding window as the original feature vector, which includes mean, variance and slope parameters; The radial basis function kernel is selected as the mapping function to calculate the similarity measure between the original feature vector and the preset feature anchor point; Based on the similarity metric, the original feature vector is nonlinearly transformed to a high-dimensional feature space to obtain a high-dimensional feature vector; The high-dimensional feature vectors are normalized to eliminate dimensional differences and preserve the relative distance relationships between features.

10. The diabetes classification and diagnostic system according to claim 7, characterized in that, The process of using wavelet transform to filter out high-frequency noise components in the aligned and interpolated data while retaining low-frequency and trend components that conform to physiological changes includes: Select wavelet basis functions suitable for physiological signals, perform multi-resolution wavelet decomposition on the aligned and interpolated time series data, and obtain wavelet coefficients of different frequency sub-bands; Set a noise threshold, and set the coefficients in the high-frequency subband whose absolute value of wavelet coefficients is lower than the noise threshold to zero to filter out high-frequency noise; Perform inverse wavelet transform on the wavelet coefficients of the low-frequency subband to reconstruct the noise-filtered time series data; Verify the consistency of trends between the reconstructed data and the original data to ensure that the low-frequency components of physiological change patterns are fully preserved.

Images